2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the BSD-type
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
17 * Redistributions in binary form must reproduce the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer in the documentation and/or other materials provided
20 * with the distribution.
22 * Neither the name of the Network Appliance, Inc. nor the names of
23 * its contributors may be used to endorse or promote products
24 * derived from this software without specific prior written
27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 #ifndef _LINUX_SUNRPC_XPRT_RDMA_H
41 #define _LINUX_SUNRPC_XPRT_RDMA_H
43 #include <linux/wait.h> /* wait_queue_head_t, etc */
44 #include <linux/spinlock.h> /* spinlock_t, etc */
45 #include <linux/atomic.h> /* atomic_t, etc */
46 #include <linux/workqueue.h> /* struct work_struct */
48 #include <rdma/rdma_cm.h> /* RDMA connection api */
49 #include <rdma/ib_verbs.h> /* RDMA verbs api */
51 #include <linux/sunrpc/clnt.h> /* rpc_xprt */
52 #include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */
53 #include <linux/sunrpc/xprtrdma.h> /* xprt parameters */
55 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
56 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
58 #define RPCRDMA_BIND_TO (60U * HZ)
59 #define RPCRDMA_INIT_REEST_TO (5U * HZ)
60 #define RPCRDMA_MAX_REEST_TO (30U * HZ)
61 #define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ)
64 * Interface Adapter -- one per transport instance
67 const struct rpcrdma_memreg_ops
*ri_ops
;
68 struct ib_device
*ri_device
;
69 struct rdma_cm_id
*ri_id
;
71 struct completion ri_done
;
73 unsigned int ri_max_segs
;
74 unsigned int ri_max_frmr_depth
;
75 unsigned int ri_max_inline_write
;
76 unsigned int ri_max_inline_read
;
77 unsigned int ri_max_send_sges
;
78 bool ri_reminv_expected
;
79 bool ri_implicit_roundup
;
80 enum ib_mr_type ri_mrtype
;
81 struct ib_qp_attr ri_qp_attr
;
82 struct ib_qp_init_attr ri_qp_init_attr
;
86 * RDMA Endpoint -- one per transport instance
93 struct ib_qp_init_attr rep_attr
;
94 wait_queue_head_t rep_connect_wait
;
95 struct rpcrdma_connect_private rep_cm_private
;
96 struct rdma_conn_param rep_remote_cma
;
97 struct sockaddr_storage rep_remote_addr
;
98 struct delayed_work rep_connect_worker
;
102 rpcrdma_init_cqcount(struct rpcrdma_ep
*ep
, int count
)
104 atomic_set(&ep
->rep_cqcount
, ep
->rep_cqinit
- count
);
107 /* To update send queue accounting, provider must take a
108 * send completion every now and then.
111 rpcrdma_set_signaled(struct rpcrdma_ep
*ep
, struct ib_send_wr
*send_wr
)
113 send_wr
->send_flags
= 0;
114 if (unlikely(atomic_sub_return(1, &ep
->rep_cqcount
) <= 0)) {
115 rpcrdma_init_cqcount(ep
, 0);
116 send_wr
->send_flags
= IB_SEND_SIGNALED
;
120 /* Pre-allocate extra Work Requests for handling backward receives
121 * and sends. This is a fixed value because the Work Queues are
122 * allocated when the forward channel is set up.
124 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
125 #define RPCRDMA_BACKWARD_WRS (8)
127 #define RPCRDMA_BACKWARD_WRS (0)
130 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
132 * The below structure appears at the front of a large region of kmalloc'd
133 * memory, which always starts on a good alignment boundary.
136 struct rpcrdma_regbuf
{
137 struct ib_sge rg_iov
;
138 struct ib_device
*rg_device
;
139 enum dma_data_direction rg_direction
;
140 __be32 rg_base
[0] __attribute__ ((aligned(256)));
144 rdmab_addr(struct rpcrdma_regbuf
*rb
)
146 return rb
->rg_iov
.addr
;
150 rdmab_length(struct rpcrdma_regbuf
*rb
)
152 return rb
->rg_iov
.length
;
156 rdmab_lkey(struct rpcrdma_regbuf
*rb
)
158 return rb
->rg_iov
.lkey
;
161 static inline struct rpcrdma_msg
*
162 rdmab_to_msg(struct rpcrdma_regbuf
*rb
)
164 return (struct rpcrdma_msg
*)rb
->rg_base
;
167 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
169 /* To ensure a transport can always make forward progress,
170 * the number of RDMA segments allowed in header chunk lists
171 * is capped at 8. This prevents less-capable devices and
172 * memory registrations from overrunning the Send buffer
173 * while building chunk lists.
175 * Elements of the Read list take up more room than the
176 * Write list or Reply chunk. 8 read segments means the Read
177 * list (or Write list or Reply chunk) cannot consume more
180 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
182 * And the fixed part of the header is another 24 bytes.
184 * The smallest inline threshold is 1024 bytes, ensuring that
185 * at least 750 bytes are available for RPC messages.
188 RPCRDMA_MAX_HDR_SEGS
= 8,
189 RPCRDMA_HDRBUF_SIZE
= 256,
193 * struct rpcrdma_rep -- this structure encapsulates state required to recv
194 * and complete a reply, asychronously. It needs several pieces of
196 * o recv buffer (posted to provider)
197 * o ib_sge (also donated to provider)
198 * o status of reply (length, success or not)
199 * o bookkeeping state to get run by reply handler (list, etc)
201 * These are allocated during initialization, per-transport instance.
203 * N of these are associated with a transport instance, and stored in
204 * struct rpcrdma_buffer. N is the max number of outstanding requests.
208 struct ib_cqe rr_cqe
;
212 struct ib_device
*rr_device
;
213 struct rpcrdma_xprt
*rr_rxprt
;
214 struct work_struct rr_work
;
215 struct list_head rr_list
;
216 struct ib_recv_wr rr_recv_wr
;
217 struct rpcrdma_regbuf
*rr_rdmabuf
;
220 #define RPCRDMA_BAD_LEN (~0U)
223 * struct rpcrdma_mw - external memory region metadata
225 * An external memory region is any buffer or page that is registered
226 * on the fly (ie, not pre-registered).
228 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
229 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
230 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
231 * track of registration metadata while each RPC is pending.
232 * rpcrdma_deregister_external() uses this metadata to unmap and
233 * release these resources when an RPC is complete.
235 enum rpcrdma_frmr_state
{
236 FRMR_IS_INVALID
, /* ready to be used */
237 FRMR_IS_VALID
, /* in use */
238 FRMR_FLUSHED_FR
, /* flushed FASTREG WR */
239 FRMR_FLUSHED_LI
, /* flushed LOCALINV WR */
242 struct rpcrdma_frmr
{
244 struct ib_cqe fr_cqe
;
245 enum rpcrdma_frmr_state fr_state
;
246 struct completion fr_linv_done
;
248 struct ib_reg_wr fr_regwr
;
249 struct ib_send_wr fr_invwr
;
254 struct ib_fmr
*fm_mr
;
259 struct list_head mw_list
;
260 struct scatterlist
*mw_sg
;
262 enum dma_data_direction mw_dir
;
264 struct rpcrdma_fmr fmr
;
265 struct rpcrdma_frmr frmr
;
267 struct rpcrdma_xprt
*mw_xprt
;
271 struct list_head mw_all
;
275 * struct rpcrdma_req -- structure central to the request/reply sequence.
277 * N of these are associated with a transport instance, and stored in
278 * struct rpcrdma_buffer. N is the max number of outstanding requests.
280 * It includes pre-registered buffer memory for send AND recv.
281 * The recv buffer, however, is not owned by this structure, and
282 * is "donated" to the hardware when a recv is posted. When a
283 * reply is handled, the recv buffer used is given back to the
284 * struct rpcrdma_req associated with the request.
286 * In addition to the basic memory, this structure includes an array
287 * of iovs for send operations. The reason is that the iovs passed to
288 * ib_post_{send,recv} must not be modified until the work request
292 /* Maximum number of page-sized "segments" per chunk list to be
293 * registered or invalidated. Must handle a Reply chunk:
296 RPCRDMA_MAX_IOV_SEGS
= 3,
297 RPCRDMA_MAX_DATA_SEGS
= ((1 * 1024 * 1024) / PAGE_SIZE
) + 1,
298 RPCRDMA_MAX_SEGS
= RPCRDMA_MAX_DATA_SEGS
+
299 RPCRDMA_MAX_IOV_SEGS
,
302 struct rpcrdma_mr_seg
{ /* chunk descriptors */
303 u32 mr_len
; /* length of chunk or segment */
304 struct page
*mr_page
; /* owning page, if any */
305 char *mr_offset
; /* kva if no page, else offset */
308 /* The Send SGE array is provisioned to send a maximum size
310 * - RPC-over-RDMA header
311 * - xdr_buf head iovec
312 * - RPCRDMA_MAX_INLINE bytes, in pages
313 * - xdr_buf tail iovec
315 * The actual number of array elements consumed by each RPC
316 * depends on the device's max_sge limit.
319 RPCRDMA_MIN_SEND_SGES
= 3,
320 RPCRDMA_MAX_PAGE_SGES
= RPCRDMA_MAX_INLINE
>> PAGE_SHIFT
,
321 RPCRDMA_MAX_SEND_SGES
= 1 + 1 + RPCRDMA_MAX_PAGE_SGES
+ 1,
324 struct rpcrdma_buffer
;
326 struct list_head rl_free
;
327 unsigned int rl_mapped_sges
;
328 unsigned int rl_connect_cookie
;
329 struct rpcrdma_buffer
*rl_buffer
;
330 struct rpcrdma_rep
*rl_reply
;
331 struct ib_send_wr rl_send_wr
;
332 struct ib_sge rl_send_sge
[RPCRDMA_MAX_SEND_SGES
];
333 struct rpcrdma_regbuf
*rl_rdmabuf
; /* xprt header */
334 struct rpcrdma_regbuf
*rl_sendbuf
; /* rq_snd_buf */
335 struct rpcrdma_regbuf
*rl_recvbuf
; /* rq_rcv_buf */
337 struct ib_cqe rl_cqe
;
338 struct list_head rl_all
;
341 struct list_head rl_registered
; /* registered segments */
342 struct rpcrdma_mr_seg rl_segments
[RPCRDMA_MAX_SEGS
];
346 rpcrdma_set_xprtdata(struct rpc_rqst
*rqst
, struct rpcrdma_req
*req
)
348 rqst
->rq_xprtdata
= req
;
351 static inline struct rpcrdma_req
*
352 rpcr_to_rdmar(struct rpc_rqst
*rqst
)
354 return rqst
->rq_xprtdata
;
358 rpcrdma_push_mw(struct rpcrdma_mw
*mw
, struct list_head
*list
)
360 list_add_tail(&mw
->mw_list
, list
);
363 static inline struct rpcrdma_mw
*
364 rpcrdma_pop_mw(struct list_head
*list
)
366 struct rpcrdma_mw
*mw
;
368 mw
= list_first_entry(list
, struct rpcrdma_mw
, mw_list
);
369 list_del(&mw
->mw_list
);
374 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
375 * inline requests/replies, and client/server credits.
377 * One of these is associated with a transport instance
379 struct rpcrdma_buffer
{
380 spinlock_t rb_mwlock
; /* protect rb_mws list */
381 struct list_head rb_mws
;
382 struct list_head rb_all
;
385 spinlock_t rb_lock
; /* protect buf lists */
386 int rb_send_count
, rb_recv_count
;
387 struct list_head rb_send_bufs
;
388 struct list_head rb_recv_bufs
;
390 atomic_t rb_credits
; /* most recent credit grant */
392 u32 rb_bc_srv_max_requests
;
393 spinlock_t rb_reqslock
; /* protect rb_allreqs */
394 struct list_head rb_allreqs
;
396 u32 rb_bc_max_requests
;
398 spinlock_t rb_recovery_lock
; /* protect rb_stale_mrs */
399 struct list_head rb_stale_mrs
;
400 struct delayed_work rb_recovery_worker
;
401 struct delayed_work rb_refresh_worker
;
403 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
406 * Internal structure for transport instance creation. This
407 * exists primarily for modularity.
409 * This data should be set with mount options
411 struct rpcrdma_create_data_internal
{
412 struct sockaddr_storage addr
; /* RDMA server address */
413 unsigned int max_requests
; /* max requests (slots) in flight */
414 unsigned int rsize
; /* mount rsize - max read hdr+data */
415 unsigned int wsize
; /* mount wsize - max write hdr+data */
416 unsigned int inline_rsize
; /* max non-rdma read data payload */
417 unsigned int inline_wsize
; /* max non-rdma write data payload */
418 unsigned int padding
; /* non-rdma write header padding */
422 * Statistics for RPCRDMA
424 struct rpcrdma_stats
{
425 unsigned long read_chunk_count
;
426 unsigned long write_chunk_count
;
427 unsigned long reply_chunk_count
;
429 unsigned long long total_rdma_request
;
430 unsigned long long total_rdma_reply
;
432 unsigned long long pullup_copy_count
;
433 unsigned long long fixup_copy_count
;
434 unsigned long hardway_register_count
;
435 unsigned long failed_marshal_count
;
436 unsigned long bad_reply_count
;
437 unsigned long nomsg_call_count
;
438 unsigned long bcall_count
;
439 unsigned long mrs_recovered
;
440 unsigned long mrs_orphaned
;
441 unsigned long mrs_allocated
;
442 unsigned long local_inv_needed
;
446 * Per-registration mode operations
449 struct rpcrdma_memreg_ops
{
450 int (*ro_map
)(struct rpcrdma_xprt
*,
451 struct rpcrdma_mr_seg
*, int, bool,
452 struct rpcrdma_mw
**);
453 void (*ro_unmap_sync
)(struct rpcrdma_xprt
*,
454 struct rpcrdma_req
*);
455 void (*ro_unmap_safe
)(struct rpcrdma_xprt
*,
456 struct rpcrdma_req
*, bool);
457 void (*ro_recover_mr
)(struct rpcrdma_mw
*);
458 int (*ro_open
)(struct rpcrdma_ia
*,
460 struct rpcrdma_create_data_internal
*);
461 size_t (*ro_maxpages
)(struct rpcrdma_xprt
*);
462 int (*ro_init_mr
)(struct rpcrdma_ia
*,
463 struct rpcrdma_mw
*);
464 void (*ro_release_mr
)(struct rpcrdma_mw
*);
465 const char *ro_displayname
;
466 const int ro_send_w_inv_ok
;
469 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops
;
470 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops
;
473 * RPCRDMA transport -- encapsulates the structures above for
474 * integration with RPC.
476 * The contained structures are embedded, not pointers,
477 * for convenience. This structure need not be visible externally.
479 * It is allocated and initialized during mount, and released
482 struct rpcrdma_xprt
{
483 struct rpc_xprt rx_xprt
;
484 struct rpcrdma_ia rx_ia
;
485 struct rpcrdma_ep rx_ep
;
486 struct rpcrdma_buffer rx_buf
;
487 struct rpcrdma_create_data_internal rx_data
;
488 struct delayed_work rx_connect_worker
;
489 struct rpcrdma_stats rx_stats
;
492 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
493 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
495 /* Setting this to 0 ensures interoperability with early servers.
496 * Setting this to 1 enhances certain unaligned read/write performance.
497 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
498 extern int xprt_rdma_pad_optimize
;
501 * Interface Adapter calls - xprtrdma/verbs.c
503 int rpcrdma_ia_open(struct rpcrdma_xprt
*, struct sockaddr
*, int);
504 void rpcrdma_ia_close(struct rpcrdma_ia
*);
505 bool frwr_is_supported(struct rpcrdma_ia
*);
506 bool fmr_is_supported(struct rpcrdma_ia
*);
509 * Endpoint calls - xprtrdma/verbs.c
511 int rpcrdma_ep_create(struct rpcrdma_ep
*, struct rpcrdma_ia
*,
512 struct rpcrdma_create_data_internal
*);
513 void rpcrdma_ep_destroy(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
514 int rpcrdma_ep_connect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
515 void rpcrdma_conn_func(struct rpcrdma_ep
*ep
);
516 void rpcrdma_ep_disconnect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
518 int rpcrdma_ep_post(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
519 struct rpcrdma_req
*);
520 int rpcrdma_ep_post_recv(struct rpcrdma_ia
*, struct rpcrdma_rep
*);
523 * Buffer calls - xprtrdma/verbs.c
525 struct rpcrdma_req
*rpcrdma_create_req(struct rpcrdma_xprt
*);
526 struct rpcrdma_rep
*rpcrdma_create_rep(struct rpcrdma_xprt
*);
527 void rpcrdma_destroy_req(struct rpcrdma_req
*);
528 int rpcrdma_buffer_create(struct rpcrdma_xprt
*);
529 void rpcrdma_buffer_destroy(struct rpcrdma_buffer
*);
531 struct rpcrdma_mw
*rpcrdma_get_mw(struct rpcrdma_xprt
*);
532 void rpcrdma_put_mw(struct rpcrdma_xprt
*, struct rpcrdma_mw
*);
533 struct rpcrdma_req
*rpcrdma_buffer_get(struct rpcrdma_buffer
*);
534 void rpcrdma_buffer_put(struct rpcrdma_req
*);
535 void rpcrdma_recv_buffer_get(struct rpcrdma_req
*);
536 void rpcrdma_recv_buffer_put(struct rpcrdma_rep
*);
538 void rpcrdma_defer_mr_recovery(struct rpcrdma_mw
*);
540 struct rpcrdma_regbuf
*rpcrdma_alloc_regbuf(size_t, enum dma_data_direction
,
542 bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia
*, struct rpcrdma_regbuf
*);
543 void rpcrdma_free_regbuf(struct rpcrdma_regbuf
*);
546 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf
*rb
)
548 return rb
->rg_device
!= NULL
;
552 rpcrdma_dma_map_regbuf(struct rpcrdma_ia
*ia
, struct rpcrdma_regbuf
*rb
)
554 if (likely(rpcrdma_regbuf_is_mapped(rb
)))
556 return __rpcrdma_dma_map_regbuf(ia
, rb
);
559 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt
*, unsigned int);
561 int rpcrdma_alloc_wq(void);
562 void rpcrdma_destroy_wq(void);
565 * Wrappers for chunk registration, shared by read/write chunk code.
568 static inline enum dma_data_direction
569 rpcrdma_data_dir(bool writing
)
571 return writing
? DMA_FROM_DEVICE
: DMA_TO_DEVICE
;
575 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
578 enum rpcrdma_chunktype
{
586 bool rpcrdma_prepare_send_sges(struct rpcrdma_ia
*, struct rpcrdma_req
*,
587 u32
, struct xdr_buf
*, enum rpcrdma_chunktype
);
588 void rpcrdma_unmap_sges(struct rpcrdma_ia
*, struct rpcrdma_req
*);
589 int rpcrdma_marshal_req(struct rpc_rqst
*);
590 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt
*);
591 void rpcrdma_reply_handler(struct work_struct
*work
);
593 /* RPC/RDMA module init - xprtrdma/transport.c
595 extern unsigned int xprt_rdma_max_inline_read
;
596 void xprt_rdma_format_addresses(struct rpc_xprt
*xprt
, struct sockaddr
*sap
);
597 void xprt_rdma_free_addresses(struct rpc_xprt
*xprt
);
598 void rpcrdma_connect_worker(struct work_struct
*work
);
599 void xprt_rdma_print_stats(struct rpc_xprt
*xprt
, struct seq_file
*seq
);
600 int xprt_rdma_init(void);
601 void xprt_rdma_cleanup(void);
603 /* Backchannel calls - xprtrdma/backchannel.c
605 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
606 int xprt_rdma_bc_setup(struct rpc_xprt
*, unsigned int);
607 int xprt_rdma_bc_up(struct svc_serv
*, struct net
*);
608 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt
*);
609 int rpcrdma_bc_post_recv(struct rpcrdma_xprt
*, unsigned int);
610 void rpcrdma_bc_receive_call(struct rpcrdma_xprt
*, struct rpcrdma_rep
*);
611 int rpcrdma_bc_marshal_reply(struct rpc_rqst
*);
612 void xprt_rdma_bc_free_rqst(struct rpc_rqst
*);
613 void xprt_rdma_bc_destroy(struct rpc_xprt
*, unsigned int);
614 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
616 extern struct xprt_class xprt_rdma_bc
;
618 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */